Root Cause Diagnosis for Time Series Data Faults Based on Fault Isolation and Channel-Attentive Sparse Causal Networks

Monitoring data from modern industrial processes are typically recorded in the form of time series. Therefore, root cause diagnosis (RCD) for time series data faults is crucial for ensuring industrial safety. In this article, an RCD framework based on fault isolation and causal discovery is proposed. First, a fault isolation method based on least absolute shrinkage and selection operator (LASSO) is applied to select candidate root cause variables. Second, a causal discovery model called channel-attentive sparse causal network (CASCN) is proposed to extract complex causal relations among variables. CASCN captures cross-time dependencies and cross-variable dependencies of multivariate time series data through a channel-independent multihead self-attention mechanism and a multichannel temporal convolutional network (TCN), respectively. Furthermore, by imposing a sparsity-inducing penalty on specific weights to encourage specific weight sets to be zero, our CASCN learns a causal matrix between all pairs of variables. Finally, the root cause variable is identified based on the causal topology. A numerical simulation case demonstrates that the proposed model achieves an average improvement of 23.10% in true-positive rate (TPR), a 60.82% reduction in false-positive rate (FPR), and a 23.47% increase in the ${F}1$ -score, with an acceptable increase in computational cost, compared to existing mainstream methods, highlighting its superiority in causal discovery. Experiments on the Tennessee Eastman process (TEP) further validate the effectiveness of the proposed RCD framework.